HCHS-Net: Un marco multimodal de características artesanales y metadatos para la clasificación interpretable de...

Accurate and timely classification of skin lesions is critical for early cancer detection, yet current deep learning approaches suffer from high computational costs, limited interpretability, and poor transparency for clinical deployment. This study presents HCHS-Net, a lightweight and interpretable multimodal framework for six-class skin lesion classification on the PAD-UFES-20 dataset. The proposed framework extracts a 116-dimensional visual feature vector through three complementary handcrafted modules: a Color Module employing multi-channel histogram analysis to capture chromatic diagnostic patterns, a Haralick Module deriving texture descriptors from the gray-level co-occurrence matrix (GLCM) that quantify surface characteristics correlated with malignancy, and a Shape Module encoding morphological properties via Hu moment invariants aligned with the clinical ABCD rule. The architectural design of HCHS-Net adopts a biomimetic approach by emulating the hierarchical information processing of the human visual system and the cognitive diagnostic workflows of expert dermatologists. Unlike conventional black-box deep learning models, this framework employs parallel processing branches that simulate the selective attention mechanisms of the human eye by focusing on biologically significant visual cues such as chromatic variance, textural entropy, and morphological asymmetry. These visual features are concatenated with a 12-dimensional clinical metadata vector encompassing patient demographics and lesion characteristics, yielding a compact 128-dimensional multimodal representation. Classification is performed through an ensemble of three gradient boosting algorithms (XGBoost, LightGBM, CatBoost) with majority voting. HCHS-Net achieves 97.76% classification accuracy with only 0.25 M parameters, outperforming deep learning baselines, including VGG-16 (94.60%), ResNet-50 (94.80%), and EfficientNet-B2 (95.16%), which require 60-97× more parameters. The framework delivers an inference time of 0.11 ms per image, enabling real-time classification on standard CPUs without GPU acceleration. Ablation analysis confirms the complementary contribution of each feature module, with metadata integration providing a 2.53% accuracy gain. The model achieves perfect melanoma and nevus recall (100%) with 99.55% specificity, maintaining reliable discrimination at safety-critical diagnostic boundaries. Comprehensive benchmarking against 13 published methods demonstrates that domain-informed handcrafted features combined with clinical metadata can match or exceed deep learning fusion approaches while offering superior interpretability and computational efficiency for point-of-care deployment.